Fuel3D is developing BioVolume and claims financial competing interests on the product. The data presented here is generated as a preproduction test on the technology. There are not specific patents granted or filed for this technology or any part of it. AstraZeneca and Medimmune do not claim any financial interests. This does not alter adherence to PLOS policies on sharing data.
Current address: Analytics Department, WundermanThompson, London, England, United Kingdom
Current address: Institute for Mathematical Innovation, University of Bath, Bath, England, United Kingdom
Current address: DENSO, Lindau (Bodensee), Germany
Current address: EMBL, Hinxton, Cambridge, England, United Kingdom
Current address: Shepherd, London, England, United Kingdom
In oncological drug development, animal studies continue to play a central role in which the volume of subcutaneous tumours is monitored to assess the efficacy of new drugs. The tumour volume is estimated by taking the volume to be that of a regular spheroid with the same dimensions. However, this method is subjective, insufficiently traceable, and is subject to error in the accuracy of volume estimates as tumours are frequently irregular.
This paper reviews the standard technique for tumour volume assessment, calliper measurements, by conducting a statistical review of a large dataset consisting of 2,500 tumour volume measurements from 1,600 mice by multiple operators across 6 mouse strains and 20 tumour models. Additionally, we explore the impact of six different tumour morphologies on volume estimation and the detection of treatment effects using a computational tumour growth model. Finally, we propose an alternative method to callipers for estimating volume–BioVolume^{TM}, a 3D scanning technique. BioVolume simultaneously captures both stereo RGB (Red, Green and Blue) images from different light sources and infrared thermal images of the tumour in under a second. It then detects the tumour region automatically and estimates the tumour volume in under a minute. Furthermore, images can be processed in parallel within the cloud and so the time required to process multiple images is similar to that required for a single image. We present data of a preproduction unit test consisting of 297 scans from over 120 mice collected by four different operators.
This work demonstrates that it is possible to record tumour measurements in a rapid minimally invasive, morphologyindependent way, and with less humanbias compared to callipers, whilst also improving data traceability. Furthermore, the images collected by BioVolume may be useful, for example, as a source of biomarkers for animal welfare and secondary drug toxicity / efficacy.
Animal models of human cancers are fundamental to our understanding of tumour biology. Tumour volume is a significant metric for preclinical trials where it provides a surrogate measure of both disease progression and treatment efficacy. Thus, accurate and repeatable estimation of tumour volume is crucial to declare a given trial to be a success or failure with confidence [
In what follows, we explore how manual calliper measurements introduce human bias into preclinical trials. Furthermore, we use a cellular automaton model to investigate how calliper measurements influence our understanding of study outcomes for six different tumour morphologies [
We analysed two datasets:


Five different strains of mice were sourced from Charles River UK (
BioVolume is a small desktop device (27 x 18.5 x 16.8 cm), which captures both thermal (infrared) and 3D surface images. To acquire a scan, the shaven mouse is held such that the tumour region is exposed to the device aperture (see
Complete set up of the BioVolume unit including computer monitor and desktop device (left), closeup image of a white SCID mouse being presented to the aperture of BioVolume (right).
The BioVolume unit consists of a stereo system with two RGB cameras, three white light flashes, and an infrared thermal camera. Upon activation, the unit collects 6 photographic (RGB) images and a single thermal frame (see
We compare two formulae (1,2) for the estimation of tumour volume:
 Spheroid formula (BioVolume & callipers):
 Cylindrical volume (BioVolume):
The cellular automaton model consisted of a rulebased model operating on two simulated cell populations growing on a 3D lattice. The rules and parametrisation originate from logical assumptions for tumour growth and treatment. There are four main parameters: vertical bias, cell division rate, magnitude, and length of treatment. These parameters are depicted in mathematical notation as:
For the calliper data set, we focus on metrics for interoperator repeatability and the consistency of BioVolume’s linear measurements with those of callipers. For the former, we use the coefficient of variation (CV) as a measure of precision and intraclass correlation (ICC) as a measure of reliability. The ICC provides a measure of the correlation between two different people acquiring scans from BioVolume on different occasions [
A full description of the methods for the analysis of the calliper and BioVolume data can be found in
Interoperator repeatability is a significant challenge when measuring subcutaneous tumours (see
Precision single values ordered by tumour model and mouse strain (A). Quantification of values within a precision limit of 0.2 (B). ICC values vs number of operators (C). Values printed on the plot indicate number of observations, dots are average ICC and shaded bars are 95% confidence intervals.
Assuming that tumour density is
BlandAltmann plot (A), linear fit with 95% confidence intervals. Proportion of mice at different levels of relative errors (B, n = 440).
To assess the impact of using calliper measurements to estimate tumour volume we developed a Cellular Automaton (CA) model of tumour growth and its treatment (see
The top row shows reconstructions of real tumours produced using BioVolume. The bottom row shows snapshots of the corresponding in silico tumours generated with the CA model. ac) depict tumours with one, two, and three peaks respectively. d) shows an iglooshaped tumour. Such tumours are characterised by a main cancerous mass (typically resembling a single peak tumour) and a “tail” and can arise if the inoculating needle leaves a trail of cells when it is retracted. e) “birthday cake” tumours can be triggered by a mutation which creates a more aggressive subpopulation of cells. f) volcanoshaped tumours can arise due to ulceration.
We simulated the growth of both control and treated tumours for each morphology described in
For each morphology, 10^{3} growth curves were generated. The grey area corresponds to the period in which the anticancer treatment was applied; I.e. days 15 to 25.
To determine the impact of simulated calliper measurements on the accuracy of treatment efficacy we calculated and compared two commonly used treatment efficacy metrics using the GT and SCderived volumes. Specifically, we computed the Tumour Growth Inhibition (TGI) and Area Under the Curve (AUC) indices (see
The TGI was computed using days 18, 24 and 30 as experiment endpoints.
We quantified the consistency between the linear length and width measurements made using BioVolume and callipers by making contemporary measurements of a given tumour using both methods and then, for each tumour, counting the number of scan measurements which fell within +/ 3mm of the calliper measurements made on the same day. These counts are displayed as histograms for both length (
There are large discrepancies in the volume–weight correlations. Firstly, the scan ellipsoid formula (using length, width and height) shows both the smallest mean discrepancy (mean: 55 mm^{3}, median: 24 mm^{3}), the smallest bias (m = 0.499) and the best correlation coefficient (R ^{2} = 0.50, see
Calliper volume was calculated using the spheroid formula, whereas scan volume corresponds to the ellipsoid volume. The linear fit is represented by the solid coloured lines, whereas the horizontal gray line is the 0 reference line. The boxplots display the median discrepancies. The hinges of each box show the 95% confidence intervals and the whiskers extend to 1.5 times the IQR.
We computed the interoperator CV for the cylindrical and spheroid volume estimates derived from BioVolume’s measurements as well as for the calliper volume estimates from Dataset 1 (calliper statistical review) and Dataset 2 (BioVolume evaluation), see
Volume estimates for BioVolume correspond to spheroid (the same formula as that used for callipers, in yellow) and cylindrical approximations (in red). Calliper data (in blue) is split into values from the BioVolume evaluation (left) and values from the Calliper statistical review (right, also in
In the calliper statistical review, we demonstrated that callipers are subject to high interoperator variability, with values reaching 130% in interoperator CV. Additionally, correlation between calliperestimated volumes and excised tumour weight was poor (
We selected a cellular automaton model to evaluate morphology since rules are easy to formulate and interpret as well as it being amenable to produce multiple morphologies with a degree of stochasticity. This is particularly interesting since we are aiming to characterise the span of multiple tumour shapes. Our computational model clearly demonstrates that assuming tumours to be spheroid as when making measurements with callipers is inadequate, particularly when tumours exhibit irregular morphologies. First, simulated calliper measurements failed to capture the decrease in tumour volume in response to treatment (
Using the prototype BioVolume scanner we were able to replicate calliper length and width measurements to within +/ 3mm in around 90% of cases (see
Finally, the interoperator variability of BioVolume outperforms that of callipers when using the spheroid formula for volume estimation. When tumour height is introduced (via the cylindrical formula, see
BioVolume, in its current form, presents a promising alternative to callipers. It has the scope to provide accurate measurements with reduced human bias. Furthermore, measurements are traceable and calibrated, as images can be revisited at any point postcapture and measurements extracted manually if required. Images can also be inspected by other users who are logged onto the system remotely potentially easing communication, peerreview, and crossvalidation. Additional work is underway to improve BioVolume’s performance and to expand its functionality. For example, machine learning can be applied to classify and characterise the stored tumour images, potentially offering additional biomarkers for treatment efficacy/toxicity and for animal welfare.
In conclusion, the use of linear calliper measurements for tumour volume estimation in lab animals is subject to significant accuracy and reproducibility problems which negatively affect the power of preclinical studies and animal welfare. We proposed BioVolume as an alternative to callipers which provides noninvasive, traceable, and more reproducible measurements with the potential to be fully morphologyindependent and to surpass callipers’ performance.
(DOCX)
(DOCX)
(CSV)
(CSV)
(CSV)
(7Z)
Thank you to AstraZeneca Bioscience Senior Management for funding this project and to all the Personal Licence holders involved in the testing. Specifically, we thank Helen Musgrove, Rebecca Whitely, Nick Moore, Emily Brough, David Simpson, Brandon Willis, Kristen Bell, Judit EspanaAgusti, Donna Goldsteen, Jane Kendrew, Graeme Smith, Chris Traher and Elizabeth Hardaker. Also, we thank the great input from Fuel3D technologies and their people to develop this technology. Special thanks to Tanya Randall, Michael Griffiths, Phalene Gowling, Fran Molina, and Chris Kane.